openenv/atari_env
3
1#!/bin/bash2# Comprehensive Docker test for Atari environment3# Tests: Build, Start, Health, Reset, Step, State, Cleanup4 5set -e # Exit on error6 7# Colors for output8RED='\033[0;31m'9GREEN='\033[0;32m'10YELLOW='\033[1;33m'11BLUE='\033[0;34m'12NC='\033[0m' # No Color13 14# Configuration15IMAGE_NAME="atari-env"16IMAGE_TAG="test"17CONTAINER_NAME="atari-env-test"18PORT="8765" # Use non-standard port to avoid conflicts19HEALTH_RETRIES=3020HEALTH_DELAY=221 22# Cleanup function23cleanup() {24 echo -e "\n${BLUE}Cleaning up...${NC}"25 docker stop ${CONTAINER_NAME} 2>/dev/null || true26 docker rm ${CONTAINER_NAME} 2>/dev/null || true27 echo -e "${GREEN}✓${NC} Cleanup complete"28}29 30# Set trap to cleanup on exit31trap cleanup EXIT32 33# Header34echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"35echo " ATARI ENVIRONMENT DOCKER TEST"36echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"37echo ""38 39# Check prerequisites40echo -e "${BLUE}Checking prerequisites...${NC}"41if ! command -v docker &> /dev/null; then42 echo -e "${RED}✗${NC} Docker is not installed"43 exit 144fi45echo -e "${GREEN}✓${NC} Docker is installed"46 47if ! command -v curl &> /dev/null; then48 echo -e "${RED}✗${NC} curl is not installed"49 exit 150fi51echo -e "${GREEN}✓${NC} curl is installed"52 53# Check if we're in the right directory54if [ ! -f "envs/atari_env/server/Dockerfile" ]; then55 echo -e "${RED}✗${NC} Must run from OpenEnv root directory"56 exit 157fi58echo -e "${GREEN}✓${NC} In correct directory"59 60# Step 1: Build Docker image61echo ""62echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"63echo -e "${BLUE}STEP 1: Building Docker Image${NC}"64echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"65 66echo "Building ${IMAGE_NAME}:${IMAGE_TAG}..."67if docker build -f envs/atari_env/server/Dockerfile -t ${IMAGE_NAME}:${IMAGE_TAG} . 2>&1 | tee /tmp/atari_build.log | tail -n 20; then68 echo -e "${GREEN}✓${NC} Docker image built successfully"69else70 echo -e "${RED}✗${NC} Docker build failed"71 echo "See /tmp/atari_build.log for full output"72 exit 173fi74 75# Check image exists76if docker image inspect ${IMAGE_NAME}:${IMAGE_TAG} &> /dev/null; then77 IMAGE_SIZE=$(docker image inspect ${IMAGE_NAME}:${IMAGE_TAG} --format='{{.Size}}' | awk '{print $1/1024/1024}')78 echo -e "${GREEN}✓${NC} Image size: ${IMAGE_SIZE} MB"79else80 echo -e "${RED}✗${NC} Image not found after build"81 exit 182fi83 84# Step 2: Start container85echo ""86echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"87echo -e "${BLUE}STEP 2: Starting Container${NC}"88echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"89 90# Clean up any existing container91docker rm -f ${CONTAINER_NAME} 2>/dev/null || true92 93echo "Starting container on port ${PORT}..."94docker run -d \95 --name ${CONTAINER_NAME} \96 -p ${PORT}:8000 \97 -e ATARI_GAME=pong \98 -e ATARI_OBS_TYPE=ram \99 -e ATARI_FRAMESKIP=4 \100 ${IMAGE_NAME}:${IMAGE_TAG}101 102if [ $? -eq 0 ]; then103 echo -e "${GREEN}✓${NC} Container started: ${CONTAINER_NAME}"104else105 echo -e "${RED}✗${NC} Failed to start container"106 exit 1107fi108 109# Wait for container to be running110sleep 2111if docker ps | grep -q ${CONTAINER_NAME}; then112 echo -e "${GREEN}✓${NC} Container is running"113else114 echo -e "${RED}✗${NC} Container is not running"115 docker logs ${CONTAINER_NAME}116 exit 1117fi118 119# Step 3: Wait for health check120echo ""121echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"122echo -e "${BLUE}STEP 3: Waiting for Server${NC}"123echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"124 125echo "Waiting for server to be ready (timeout: ${HEALTH_RETRIES}s)..."126for i in $(seq 1 ${HEALTH_RETRIES}); do127 if curl -s http://localhost:${PORT}/health > /dev/null 2>&1; then128 echo -e "${GREEN}✓${NC} Server is ready (${i}s)"129 break130 fi131 132 if [ $i -eq ${HEALTH_RETRIES} ]; then133 echo -e "${RED}✗${NC} Server did not become ready in time"134 echo "Container logs:"135 docker logs ${CONTAINER_NAME}136 exit 1137 fi138 139 echo -n "."140 sleep ${HEALTH_DELAY}141done142 143# Step 4: Test health endpoint144echo ""145echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"146echo -e "${BLUE}STEP 4: Testing Health Endpoint${NC}"147echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"148 149HEALTH_RESPONSE=$(curl -s http://localhost:${PORT}/health)150echo "Response: ${HEALTH_RESPONSE}"151 152if echo "${HEALTH_RESPONSE}" | grep -q "healthy"; then153 echo -e "${GREEN}✓${NC} Health endpoint working"154else155 echo -e "${RED}✗${NC} Health endpoint failed"156 exit 1157fi158 159# Step 5: Test reset endpoint160echo ""161echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"162echo -e "${BLUE}STEP 5: Testing Reset Endpoint${NC}"163echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"164 165RESET_RESPONSE=$(curl -s -X POST http://localhost:${PORT}/reset -H "Content-Type: application/json" -d '{}')166 167if [ -z "${RESET_RESPONSE}" ]; then168 echo -e "${RED}✗${NC} Reset endpoint returned empty response"169 docker logs ${CONTAINER_NAME} | tail -20170 exit 1171fi172 173echo "Response (first 200 chars): ${RESET_RESPONSE:0:200}..."174 175# Check if response contains expected fields176if echo "${RESET_RESPONSE}" | grep -q "observation" && \177 echo "${RESET_RESPONSE}" | grep -q "screen" && \178 echo "${RESET_RESPONSE}" | grep -q "legal_actions"; then179 echo -e "${GREEN}✓${NC} Reset endpoint working"180 181 # Extract some info182 SCREEN_LEN=$(echo "${RESET_RESPONSE}" | grep -o '"screen":\[[^]]*\]' | wc -c)183 echo " Screen data length: ${SCREEN_LEN} chars"184else185 echo -e "${RED}✗${NC} Reset response missing required fields"186 echo "Full response: ${RESET_RESPONSE}"187 exit 1188fi189 190# Step 6: Test step endpoint191echo ""192echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"193echo -e "${BLUE}STEP 6: Testing Step Endpoint${NC}"194echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"195 196STEP_PAYLOAD='{"action": {"action_id": 0, "game_name": "pong"}}'197STEP_RESPONSE=$(curl -s -X POST http://localhost:${PORT}/step -H "Content-Type: application/json" -d "${STEP_PAYLOAD}")198 199if [ -z "${STEP_RESPONSE}" ]; then200 echo -e "${RED}✗${NC} Step endpoint returned empty response"201 docker logs ${CONTAINER_NAME} | tail -20202 exit 1203fi204 205echo "Response (first 200 chars): ${STEP_RESPONSE:0:200}..."206 207# Check if response contains expected fields208if echo "${STEP_RESPONSE}" | grep -q "observation" && \209 echo "${STEP_RESPONSE}" | grep -q "reward" && \210 echo "${STEP_RESPONSE}" | grep -q "done"; then211 echo -e "${GREEN}✓${NC} Step endpoint working"212 213 # Extract reward and done214 REWARD=$(echo "${STEP_RESPONSE}" | grep -o '"reward":[^,}]*' | cut -d: -f2)215 DONE=$(echo "${STEP_RESPONSE}" | grep -o '"done":[^,}]*' | cut -d: -f2)216 echo " Reward: ${REWARD}"217 echo " Done: ${DONE}"218else219 echo -e "${RED}✗${NC} Step response missing required fields"220 echo "Full response: ${STEP_RESPONSE}"221 exit 1222fi223 224# Step 7: Test state endpoint225echo ""226echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"227echo -e "${BLUE}STEP 7: Testing State Endpoint${NC}"228echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"229 230STATE_RESPONSE=$(curl -s http://localhost:${PORT}/state)231 232if [ -z "${STATE_RESPONSE}" ]; then233 echo -e "${RED}✗${NC} State endpoint returned empty response"234 docker logs ${CONTAINER_NAME} | tail -20235 exit 1236fi237 238echo "Response: ${STATE_RESPONSE}"239 240# Check if response contains expected fields241if echo "${STATE_RESPONSE}" | grep -q "episode_id" && \242 echo "${STATE_RESPONSE}" | grep -q "step_count" && \243 echo "${STATE_RESPONSE}" | grep -q "game_name"; then244 echo -e "${GREEN}✓${NC} State endpoint working"245 246 # Extract info247 GAME_NAME=$(echo "${STATE_RESPONSE}" | grep -o '"game_name":"[^"]*"' | cut -d'"' -f4)248 STEP_COUNT=$(echo "${STATE_RESPONSE}" | grep -o '"step_count":[^,}]*' | cut -d: -f2)249 echo " Game: ${GAME_NAME}"250 echo " Steps: ${STEP_COUNT}"251else252 echo -e "${RED}✗${NC} State response missing required fields"253 echo "Full response: ${STATE_RESPONSE}"254 exit 1255fi256 257# Step 8: Test multiple steps258echo ""259echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"260echo -e "${BLUE}STEP 8: Testing Multiple Steps${NC}"261echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"262 263echo "Taking 10 steps..."264TOTAL_REWARD=0265for i in {1..10}; do266 ACTION_ID=$((RANDOM % 3)) # Random action 0-2267 STEP_PAYLOAD="{\"action\": {\"action_id\": ${ACTION_ID}, \"game_name\": \"pong\"}}"268 STEP_RESPONSE=$(curl -s -X POST http://localhost:${PORT}/step -H "Content-Type: application/json" -d "${STEP_PAYLOAD}")269 270 if ! echo "${STEP_RESPONSE}" | grep -q "observation"; then271 echo -e "${RED}✗${NC} Step ${i} failed"272 exit 1273 fi274 275 REWARD=$(echo "${STEP_RESPONSE}" | grep -o '"reward":[^,}]*' | cut -d: -f2 | sed 's/null/0/')276 DONE=$(echo "${STEP_RESPONSE}" | grep -o '"done":[^,}]*' | cut -d: -f2)277 278 echo " Step ${i}: action=${ACTION_ID}, reward=${REWARD}, done=${DONE}"279 280 if [ "${DONE}" = "true" ]; then281 echo " Episode completed early at step ${i}"282 break283 fi284done285 286echo -e "${GREEN}✓${NC} Multiple steps completed successfully"287 288# Step 9: Check container logs for errors289echo ""290echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"291echo -e "${BLUE}STEP 9: Checking Container Logs${NC}"292echo -e "${BLUE}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${NC}"293 294LOGS=$(docker logs ${CONTAINER_NAME} 2>&1)295 296if echo "${LOGS}" | grep -i "error" | grep -v "LoggerMode.Error"; then297 echo -e "${YELLOW}⚠${NC} Found errors in logs:"298 echo "${LOGS}" | grep -i "error" | head -5299else300 echo -e "${GREEN}✓${NC} No errors in container logs"301fi302 303if echo "${LOGS}" | grep -i "exception"; then304 echo -e "${RED}✗${NC} Found exceptions in logs:"305 echo "${LOGS}" | grep -i "exception" | head -5306 exit 1307else308 echo -e "${GREEN}✓${NC} No exceptions in container logs"309fi310 311# Final Summary312echo ""313echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"314echo -e "${GREEN}✅ ALL DOCKER TESTS PASSED${NC}"315echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"316echo ""317echo "Summary:"318echo " ✓ Docker image built successfully"319echo " ✓ Container started and ran"320echo " ✓ Health endpoint working"321echo " ✓ Reset endpoint working"322echo " ✓ Step endpoint working"323echo " ✓ State endpoint working"324echo " ✓ Multiple steps working"325echo " ✓ No errors or exceptions"326echo ""327echo "Image: ${IMAGE_NAME}:${IMAGE_TAG}"328echo "Container: ${CONTAINER_NAME}"329echo "Port: ${PORT}"330echo ""331echo "To keep container running: docker start ${CONTAINER_NAME}"332echo "To view logs: docker logs ${CONTAINER_NAME}"333echo ""334